Generalized Linear Models Linear Models: Expected value is additive (“linear”) combination of parameters Generalized Linear Models: Expected value is some function of an additive combination of parameters Y i ∼ Normal(μ i , σ) μ i = α + β X X i + β Z Z i Y i ∼ Bernoulli(p i ) f(p i ) = α + β X X i + β Z Z i f(p i )
logit(p i ) = α + βx i Generalized Linear Models: Expected value is some function of an additive combination of parameters Uniform changes in predictor not uniform changes in prediction All predictor variables interact, moderate one another In uences predictions & uncertainty of predictions
G D A u set.seed(12) N # even gender distribution G # sample ability, high (1) to average (0) u # gender 1 tends to apply to department 1, 2 to 2 # and G=1 with greater ability tend to apply to 2 as well D # matrix of acceptance rates [dept,gender] p_u0 p_u1 p_u # simulate acceptance p A
set.seed(12) N # even gender distribution G # sample ability, high (1) to average (0) u # gender 1 tends to apply to department 1, 2 to 2 # and G=1 with greater ability tend to apply to 2 as well D # matrix of acceptance rates [dept,gender] p_u0 p_u1 p_u # simulate acceptance p A G D A u
set.seed(12) N # even gender distribution G # sample ability, high (1) to average (0) u # gender 1 tends to apply to department 1, 2 to 2 # and G=1 with greater ability tend to apply to 2 as well D # matrix of acceptance rates [dept,gender] p_u0 p_u1 p_u # simulate acceptance p A G D A u
set.seed(12) N # even gender distribution G # sample ability, high (1) to average (0) u # gender 1 tends to apply to department 1, 2 to 2 # and G=1 with greater ability tend to apply to 2 as well D # matrix of acceptance rates [dept,gender] p_u0 p_u1 p_u # simulate acceptance p A G D A u > table(G,D,u) , , u = 0 D G 1 2 1 908 0 2 220 645 , , u = 1 D G 1 2 1 0 114 2 28 85
set.seed(12) N # even gender distribution G # sample ability, high (1) to average (0) u # gender 1 tends to apply to department 1, 2 to 2 # and G=1 with greater ability tend to apply to 2 as well D # matrix of acceptance rates [dept,gender] p_u0 p_u1 p_u # simulate acceptance p A G D A u > p_u0 [,1] [,2] [1,] 0.1 0.1 [2,] 0.1 0.3 > p_u1 [,1] [,2] [1,] 0.3 0.5 [2,] 0.3 0.5 Department 2 discriminates against gender 1
You guessed it: Collider bias Stratifying by D opens non- causal path through u Can estimate total causal e ect of G, but isn’t what we want Cannot estimate direct e ect of D or G Worst place in the world, Collider Bias country
You guessed it: Collider bias More intuitive explanation: High ability G1s apply to discriminatory department anyway G1s in that department are higher ability on average than G2s High ability compensates for discrimination => masks evidence G D A u
Citation networks Citation networks of members of NAS Women associated with lower lifetime citation rate G C gender citations Lerman et al 2022 Gendered citation patterns among the scienti c elite
Membership Elections to NAS Women associated with 3–15 times higher election rate, controlling for citations gender Card et al 2022 Gender Gaps at the Academies NAS member citations G M C
gender NAS member citations G M C q quality (hidden) Restrict sample to NAS members, examine citations If men less likely to be elected, then must have higher q,C to compensate
gender NAS member citations G M C q quality (hidden) Control for citations, examine elections to NAS G is “treatment”. C is a post-treatment variable! If women less likely to be cited (bias), then women more likely to be elected because they have higher q than indicated by C
No causes in; No causes out Hyped papers with vague estimands, unwise adjustment sets Policy design through collider bias? We can do better Strong assumptions required Qualitative data useful G M C q
Sensitivity analysis What are the implications of what we don’t know? Assume confound exists, model its consequences for di erent strengths/kinds of in uence How strong must the confound be to change conclusions? G D A u
Sensitivity analysis A i ∼ Bernoulli(p i ) logit(p i ) = α[G i , D i ] + β G[i] u i G D A u (D i = 2) ∼ Bernoulli(q i ) logit(q i ) = δ[G i ] + γ G[i] u i G D A u
A i ∼ Bernoulli(p i ) logit(p i ) = α[G i , D i ] + β G[i] u i (D i = 2) ∼ Bernoulli(q i ) logit(q i ) = δ[G i ] + γ G[i] u i datl datl$D2 datl$N datl$b datl$g mGDu alist( # A model A ~ bernoulli(p), logit(p) matrix[G,D]:a ~ normal(0,1), # D model D2 ~ bernoulli(q), logit(q) delta[G] ~ normal(0,1), # declare unobserved u vector[N]:u ~ normal(0,1) ), data=datl , chains=4 , cores=4 ) u j ∼ Normal(0,1)
Sensitivity analysis What are the implications of what we don’t know? Somewhere between pure simulation and pure analysis Vary confound strength over range and show how results change –or– vary other e ects and estimate confound strength Confounds persist — don’t pretend G D A u 0.0 0.1 0.2 0.3 0.4 0.5 0 10 20 30 40 probability admission Density
More parameters (2006) than observations (2000)! A i ∼ Bernoulli(p i ) logit(p i ) = α[G i , D i ] + β G[i] u i (D i = 2) ∼ Bernoulli(q i ) logit(q i ) = δ[G i ] + γ G[i] u i u j ∼ Normal(0,1)
Oceanic Technology How is technological complexity related to population size? To social structure? data(Kline) In uence of population size and contact on total tools Kline & Boyd 2010 Population size predicts technological complexity in Oceania
Modeling tools Tool count is not binomial: No maximum Poisson distribution: Very high maximum and very low probability of each success Here: Many many possible technologies, very few realized in any one place
Poisson link is log Poisson distribution takes shape from expected value Must be positive Exponential scaling can be surprising! Y i ∼ Poisson(λ i ) log(λ i ) = α + βx i λ i = exp(α + βx i ) log(λ i ) exp
Poisson (poison) priors Exponential scaling can be surprising α ∼ Normal(0,10) log(λ i ) = α 0 20 40 60 80 100 0.00 0.01 0.02 0.03 0.04 0.05 mean number of tools Density
> compare( m11.9 , m11.10 , func=PSIS ) Some Pareto k values are high (>0.5). Set pointwise=TRUE to inspect individual points. Some Pareto k values are high (>0.5). Set pointwise=TRUE to inspect individual points. PSIS SE dPSIS dSE pPSIS weight m11.10 85.9 13.50 0.0 NA 7.3 1 m11.9 141.3 33.69 55.4 33.13 8.0 0 pPSIS is the “penalty”, the e ective number of parameters. How can a model with more parameters (m11.10) have fewer e ective parameters? m11.9 changes more when individual societies are dropped
is model is wack Two immediate ways to improve the model (1) Use a robust model: gamma-Poisson (neg-binomial) (2) A principled scienti c model 0 50000 150000 250000 0 20 40 60 population total tools Tonga Hawaii
ΔT = α C Pβ C − γT 0 10 20 30 40 50 0 2 4 6 8 10 time Tools P = 1e4 P = 1e3 f T for ( i in 2:t_max ) T[i] return(T) } plot( NULL , xlim=c(0,50) , ylim=c(0,10) , xlab="time" , ylab="Tools" ) T lines( 1:50 , T , lwd=3 , col=2 ) T lines( 1:50 , T , lwd=3 , col=2 )
# innovation/loss model dat2 C=d$contact_id ) m11.11 alist( T ~ dpois( lambda ), lambda a[C] ~ dnorm(1,1), b[C] ~ dexp(1), g ~ dexp(1) ), data=dat2 , chains=4 , cores=4 ) All parameters must be positive Two ways to do this (1) use exp() (2) use appropriate prior ̂ T = α C Pβ C γ
0 50000 150000 250000 0 20 40 60 population total tools Tonga Hawaii 0 50000 150000 250000 0 20 40 60 population total tools Tonga Hawaii Innovation/loss model Generalized linear model Still have to deal with location as confound
Count GLMs Distributions from constraints Maximum entropy priors: Binomial, Poisson, and extensions More event types: Multinomial and categorical Robust regressions: Beta-binomial, gamma-Poisson
Simpson’s Pandora’s Box Simpson’s paradox: Reversal of an association when groups are combined or separated No way to know which association (combined/separated) is correct without causal assumptions In nite evils unleashed Simpson 1951. e Interpretation of Interaction in Contingency Tables The container HELD every imaginable estimate. Pandora opens a spreadsheet Simpson’s paradox plagues us to this day.
Berkeley Paradox Unconditional on department: Women admitted at lower rate Conditional on department: Women admitted slightly more Which is correct? No way to know without assumptions dept admit reject applications gender 1 A 512 313 825 male 2 A 89 19 108 female 3 B 353 207 560 male 4 B 17 8 25 female 5 C 120 205 325 male 6 C 202 391 593 female 7 D 138 279 417 male 8 D 131 244 375 female 9 E 53 138 191 male 10 E 94 299 393 female 11 F 22 351 373 male 12 F 24 317 341 female
Berkeley Paradox Which is correct? No way to know without assumptions Mediator (department) Collider + confound (ability) dept admit reject applications gender 1 A 512 313 825 male 2 A 89 19 108 female 3 B 353 207 560 male 4 B 17 8 25 female 5 C 120 205 325 male 6 C 202 391 593 female 7 D 138 279 417 male 8 D 131 244 375 female 9 E 53 138 191 male 10 E 94 299 393 female 11 F 22 351 373 male 12 F 24 317 341 female
Z = 1 Z = 0 X Z Y -3 -2 -1 0 1 2 3 4 -3 -2 -1 0 1 2 3 X Y cols N Z X Y plot( X , Y , col=cols[Z+1] , lwd=3 ) abline(lm(Y[Z==1]~X[Z==1]),col=2,lwd=3) abline(lm(Y[Z==0]~X[Z==0]),col=4,lwd=3) abline(lm(Y~X),lwd=3)
Z = 1 Z = 0 cols N X Z Y plot( X , Y , col=cols[Z+1] , lwd=3 ) abline(lm(Y[Z==1]~X[Z==1]),col=2,lwd=3) abline(lm(Y[Z==0]~X[Z==0]),col=4,lwd=3) abline(lm(Y~X),lwd=3) X Z Y -3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3 X Y
Z = 1 Z = 0 cols N X Y Z plot( X , Y , col=cols[Z+1] , lwd=3 ) abline(lm(Y[Z==1]~X[Z==1]),col=2,lwd=3) abline(lm(Y[Z==0]~X[Z==0]),col=4,lwd=3) abline(lm(Y~X),lwd=3) X Z Y -2 -1 0 1 2 3 -2 -1 0 1 2 3 X Y
# base rate differences erasing effect of X N X Z p Y mX alist( Y ~ dbern(p), logit(p) a ~ dnorm(0,1), b ~ dnorm(0,1) ) , data=list(Y=Y,X=X) ) mXZ alist( Y ~ dbern(p), logit(p) a ~ dnorm(0,1), b[Z] ~ dnorm(0,1) ) , data=list(Y=Y,X=X,Z=Z+1) ) X Y Z
X Y Z # base rate differences erasing effect of X N X Z p Y mX alist( Y ~ dbern(p), logit(p) a ~ dnorm(0,1), b ~ dnorm(0,1) ) , data=list(Y=Y,X=X) ) mXZ alist( Y ~ dbern(p), logit(p) a ~ dnorm(0,1), b[Z] ~ dnorm(0,1) ) , data=list(Y=Y,X=X,Z=Z+1) ) logit(p i ) = α + βX i
# base rate differences erasing effect of X N X Z p Y mX alist( Y ~ dbern(p), logit(p) a ~ dnorm(0,1), b ~ dnorm(0,1) ) , data=list(Y=Y,X=X) ) mXZ alist( Y ~ dbern(p), logit(p) a ~ dnorm(0,1), b[Z] ~ dnorm(0,1) ) , data=list(Y=Y,X=X,Z=Z+1) ) X Y Z logit(p i ) = α + βX i logit(p i ) = α + β Z[i] X i
Simpson’s Pandora’s Box No paradox, because almost anything can produce it People do not have intuitions about coe cient reversals Stop naming statistical paradoxes; start teaching scienti c logic Simpson 1951. e Interpretation of Interaction in Contingency Tables The container HELD every imaginable estimate. Pandora opens a spreadsheet Simpson’s paradox plagues us to this day.
rmcelreath
matleenalaakso
skimatalk
signer
fdmiyata
funain
ange
asial_edu
tkimura12
cbtlibrary
natsuyo
corocn
addyosmani
denniskardys
malarkey
tammielis
mongodb
eileencodes
philhawksworth
aleyda
paulrobertlloyd
keathley
robhawkes
sachag